🤖 AI

AI YOU Town: Make Friends and Money with Your Digital Twin

This paper introduces **AI YOU**, a framework that enhances digital twin consistency and uncertainty calibration by combining prompting, Bayesian updating, and conformal prediction to dynamically update a 22-dimensional personality profile and preserve persona fidelity over long, multi-agent interactions.

Original authors: Yan Lin, Yuyang Dai, Jiahui Geng, Yuxia Wang

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Yan Lin, Yuyang Dai, Jiahui Geng, Yuxia Wang

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a bustling virtual town called AI YOU Town, where your personal "digital twin" hangs out, makes friends, and even takes on jobs like a counselor or an investor. But here's the catch: most digital twins today are like actors who memorize a script once and then forget who they are halfway through the play. They drift, they contradict themselves, and they guess your personality without knowing how sure they should be.

The researchers behind AI YOU say, "Let's fix that." They built a system that doesn't just guess who you are; it learns who you are, updates its guess as you talk, and knows exactly when it's unsure.

The Problem: The "One-and-Done" Script

Think of old digital twins like a student who reads a biography of a famous person once and then tries to act like them forever. If the student forgets a detail or gets confused, they just make things up. The paper argues that current methods are too rigid. They rely on a single "prompt" (a set of instructions) and treat every new sentence you say as if it's the only thing that matters, ignoring the history of your conversation. This causes the twin to "drift"—suddenly acting like a grumpy old man when you were just a cheerful teenager five minutes ago.

The Solution: The "Detective with a Notebook"

AI YOU Town introduces a smarter approach. Instead of a static script, the system acts like a super-detective with a three-layer notebook:

The Working Memory (The Sticky Note): This holds the last 20 turns of conversation. It's your immediate context, like remembering what you just said in the last few minutes.
The Episodic Memory (The Photo Album): Every 10 turns, the system takes a snapshot of the conversation, compresses it into a summary, and files it away. It's like saving a photo album of your chats so the twin doesn't forget the big events.
The Semantic Memory (The Encyclopedia): Every 50 turns, the system looks through the photo albums to find deeper patterns—like your core values or long-term goals—and updates its "encyclopedia" of you.

This three-layer memory is the secret sauce. It allows the twin to stay consistent over long conversations (even 100 turns!) without getting confused or forgetting who you are.

The "22-Dimensional" Personality Map

Instead of just saying "You are nice," AI YOU builds a detailed profile with 22 different dimensions. This goes beyond the standard "Big Five" personality traits (like being outgoing or organized) to include things like how anxious you might feel, how lonely you perceive yourself to be, and even your attachment style in relationships.

But here is the most important part: The system knows when it's guessing.

The "Uncertainty Meter"

Imagine a weather app that says, "It will rain," but doesn't tell you if it's a 10% chance or a 90% chance. That's dangerous. AI YOU is different. It uses a special math trick called Bayesian updating and conformal prediction to act like a weather app with a confidence meter.

How it works: If you say something vague, the system says, "I'm not sure yet," and keeps its options open. As you talk more and provide more clues, it narrows down the guess.
The Proof: In their tests, the system achieved a "coverage" rate between 0.921 and 0.976. This means that when the system says, "I'm 90% sure this trait is X," it is actually correct about 92% to 97% of the time. That's a huge improvement over standard methods, which often pretend to be confident when they are actually just guessing.

What the System Doesn't Do

The paper is very clear about what this is not:

It is not a mind reader. It cannot know your secrets unless you tell them.
It is not a replacement for a real therapist or a professional counselor. The authors explicitly state this is a research prototype, not a tool for medical advice or hiring decisions.
It does not work perfectly for every AI model. The system works best with powerful "API" models (like GPT-5.4 or Qwen-3.6-Max). When they tried it on smaller, local models, the results were sometimes worse, suggesting that the complex math needs a strong brain to work right.

The Results: Less Drift, More Truth

When they tested this in a "Werewolf" game (a social deduction game where players try to trick each other) and in long role-playing sessions, the AI YOU twins stayed true to their personalities much better than the old static twins.

In the role-playing tests, the "Refresh" method (using the memory anchor) improved consistency scores by about 0.84 to 0.91 points on a 5-point scale compared to the old way.
In the "Werewolf" game, the system reduced the "drift" (how much the twin changed its personality) by a significant margin for most models.

The Bottom Line

AI YOU Town suggests that the future of digital twins isn't about writing a perfect script once. It's about building a living, breathing memory system that updates its understanding of you as you talk, knows when it's unsure, and keeps your personality consistent even after hours of chatting.

The authors suggest this is a big step forward, but they also warn that we need to be careful. Because these twins are so good at mimicking us, we need strong rules to make sure they don't trick us or make us feel emotionally dependent on them. For now, it's a fascinating experiment in a virtual town, waiting for real people to move in.

Technical Summary: AI YOU Town

Problem Statement

Existing approaches to inferring user traits and generating persona-consistent responses rely heavily on static prompting. These methods suffer from three primary deficiencies:

Lack of Calibrated Uncertainty: Single-pass Large Language Model (LLM) predictions produce point estimates without reliable confidence scores, failing to distinguish between strong evidence and speculation.
Absence of Sequential Updating: Most pipelines treat conversational utterances independently, lacking mechanisms to accumulate evidence and update beliefs over time.
Disconnect Between Inference and Generation: Personality inference and persona-driven generation are typically treated as separate problems, leading to a gap between understanding a user and embodying them. Furthermore, static personas tend to "drift" during long interactions, contradicting earlier statements or abandoning role-appropriate behavior.

Methodology

The authors present AI YOU, an interactive multi-agent framework that unifies personality inference, modeling, and generation. The system is prototyped as AI YOU Town, a virtual environment where personal digital twins (PDTs) interact, socialize, and work.

Core Components

22-Dimensional Personality Inference:
- The system estimates a structured profile covering 22 dimensions, extending beyond the standard Big Five to include attachment styles, self-efficacy, perceived loneliness, emotional states, and MBTI types.
- Inference Strategy: Uses Direct prompting and Chain-of-Thought (CoT) to extract candidate observations. It employs contrastive extraction to downweight attributes supported only by weak context.
- Uncertainty Quantification:
  - Bayesian Updating: For numeric traits (e.g., Big Five scores), the system uses Gaussian conjugate updates to refine posterior means and variances based on confidence scores.
  - Conformal Prediction: For categorical and set-valued traits, it generates prediction sets with coverage guarantees (targeting 90% empirical coverage) to formally represent uncertainty.
Three-Layer Cognitive Memory:
To maintain consistency over long horizons, the system utilizes a memory architecture inspired by human cognition:
- Working Memory: A FIFO buffer of the last 20 turns for immediate context.
- Episodic Memory: Compressed summaries of interaction events and emotional trends, updated every 10 turns.
- Semantic Memory: High-level user patterns and relationship insights extracted from episodic memories, updated every 50 turns.
- Memory Anchor: A periodically refreshed persona anchor (every $k$ turns) that integrates the current anchor state with recent dialogue and trait constraints to prevent drift.
Affective, Relationship, and Risk Monitors:
- Affective Monitor: Tracks emotional state and intensity to guide response strategies (e.g., validation vs. boundary setting).
- Relationship Monitor: Models interpersonal dynamics and sentiment progression.
- Risk Monitor: Detects unsafe patterns such as manipulation, romance scams, and "pig-butchering" investment schemes using rule-based matching and LLM semantic analysis.
Persona-Conditioned Generation:
The system generates responses conditioned on the refreshed memory anchor and recent context rather than the entire dialogue history, ensuring the twin remains consistent with the inferred profile.

Key Contributions

Unified Framework: The first system to organically connect personality inference, persona modeling/preservation, and persona-conditioned generation within a single pipeline.
Calibrated State Estimation: Integration of Bayesian posterior inference and conformal prediction to provide calibrated uncertainty estimates for user traits, moving beyond simple point estimates.
Long-Horizon Consistency: A three-layer memory mechanism and periodic anchor refresh strategy that significantly reduces persona drift in interactions exceeding 100 turns.
Prototype Implementation: The creation of "AI YOU Town," a bidirectional ecosystem connecting real users with their digital twins for companionship and consulting.

Experimental Results

The framework was evaluated across multiple benchmarks (PANDORA, Essays, DailyDialog, PsyScam, LoCoMo, PersonaConflicts) using various backbones (GPT-5.4, Qwen-3.6-Max, DeepSeek-V4, OLMo).

Uncertainty Calibration: AI YOU significantly improved Expected Calibration Error (ECE), reducing it by 1.5–1.7× compared to baselines. Conformal prediction sets achieved empirical coverage between 0.921 and 0.976, consistently meeting the 90% target.
Trait Estimation: While raw regression accuracy (MAE) saw modest improvements, the system's ability to represent ambiguous traits as uncertain was superior. Ablation studies confirmed that removing Bayesian aggregation or conformal calibration degraded performance.
Persona Preservation: In adversarial settings (e.g., Werewolf game) and long dialogues (100 turns), the periodic memory refresh ("Refresh") reduced trait drift compared to static prompting. For instance, on GPT-4o-mini, the mean absolute deviation from target traits dropped from 0.036 (Static) to 0.008 (Refresh).
State Monitoring: The risk monitor improved partial scores in scam detection (PsyScam) for API models, though the authors note that structured monitoring can over-constrain smaller local models. Memory ablations showed that removing memory entirely caused the most significant performance degradation in long-session QA (LoCoMo).

Significance and Claims

The paper posits that AI YOU represents a shift from static, one-time persona prompting to a longitudinal, evidence-governed state-maintenance system.

Reliability over Raw Accuracy: The primary contribution is not necessarily a massive leap in raw trait prediction accuracy, but a more controlled, calibrated way of accumulating uncertain evidence. This allows the system to distinguish between what is known and what remains uncertain, which is critical for safe and reliable persona conditioning.
Design Principle: The work suggests that future digital twin systems should be evidence-governed, uncertainty-aware, and inspectable, with generations conditioned on supported and privacy-filtered states.
Modesty: The authors explicitly state that the current prototype is a research tool. They acknowledge limitations, including the lack of end-to-end longitudinal user studies, reliance on text-based inference (which can be sparse or biased), and the use of simulated benchmarks for persona preservation. They emphasize that the system is not a substitute for professional advice or identity verification.